PRE-SPECIFIED BEFORE LOOKING AT THE OUTCOME CORRELATION (2026-08-25)

Claim under test: note 0075 asserts "one seat's concession is a TRANSFER that lifts the
worst-off; when all five concede the concessions CANCEL". That is currently a reading, not
a measurement. This tests it WITHIN A6 only.

Why not cross-arm: A5 has exactly one advised seat, so only one seat can ever have a
measurable advice_overridden_toward (confirmed: 68/68 A5 episodes have exactly 1). Any
A5-vs-A6 "dispersion of conceding" comparison is confounded by construction.

Within A6 there is real variation: episodes have 1..5 seats with a measurable override
(14/40/39/17/5), so concentration of conceding is identifiable.

Definitions:
  gave_up_i    = sum over seat i's overrides of max(0, -advice_overridden_toward)
  n_conceders  = number of seats with gave_up_i > 0
  concentration = Herfindahl index of the gave_up vector (1.0 = one seat did all the
                  conceding; 0.2 = five seats gave up equally)
  outcome      = that episode's normalized_gini (closed deals only)

Prediction if the TRANSFER reading is right:
  concentration correlates NEGATIVELY with normalized_gini
  (conceding concentrated in one seat -> flatter deal, i.e. a real transfer)
  equivalently n_conceders correlates POSITIVELY with normalized_gini.

FALSIFIER, stated in advance: if both correlations are null (CI covering zero), the
"concessions cancel" mechanism is NOT supported by the within-arm data, and note 0075 must
label it an untested conjecture rather than a reading of the evidence.

Statistics: Pearson r, cluster bootstrap over the 24 instances, 10,000 draws, seed 20260825.